Neural Networks for Fashion Image Classification and Visual Search

In modern internet commerce and digital retail, fashion classification and visual search are very important jobs. They make it possible to quickly recommend products, find them, and keep track of inventories. Even though deep learning has come a long way, current CNN (convolutional neural network) methods still have trouble with class disparities, overlapping categories, and picking up on fine-grained visual details in manner datasets. To solve these problems, this study suggests Fashion ViT-SA, a hybrid neural network that combines a Vision Transformer (ViT-Base, Patch16) core with a Spatial Attention Module. The model uses the transformer's capacity to encode global context while also using spatial attention to highlight local clothing features, which improves discriminative representation. We used the Deep Fashion multi-modal Dataset and did some preprocessing, such as filtering categories, encoding labels, separating the data into groups, and adding data both online and offline to make sure the class distributions were even. Fashion ViT-SA was used to extract features that were then used to sort items into seven fashion types and to search for items visually based on their content using an Annoy-inspired approximate nearest neighbour index. We trained the model with weighted cross-entropy loss, improved it with Adam-W, and then tested it for accuracy, precision, recall, F1-score, and loss. Experimental findings indicate that Fashion ViT-SA attains 83% accuracy, surpassing a baseline CNN model by 13%, and delivers solid, real-time retrieval performance for visually analogous products. The research underscores the promise of hybrid transformer-based architectures in fashion AI applications, merging classification precision with scalable visual search, thus propelling both scholarly inquiry and practical e-commerce innovations.